Development of a Hybrid GAN Using CGAN and LS-Wasserstein Loss for Synthetic Data Generation in COVID-19 Severity Prediction

Development of a Hybrid GAN Using CGAN and LS-Wasserstein Loss for Synthetic Data Generation in COVID-19 Severity Prediction. To overcome the limitations of scarce and imbalanced chest X-ray data, three GAN variants CGAN, LS-GAN and WGAN were combined to develop a Hybrid GAN. 809 (538 COVID-19 and 271 Non-COVID-19) chest radiography images were obtained from the COVID-19 patient radiographic images and case note from the isolation centers of Nigeria hospital and GitHub repository. The CGAN, LS-GAN, WGAN and LS+W+CGAN were implemented to generate realistic, class-conditional synthetic images. The hybrid LS+W+CGAN achieved the best balance between training stability, visual fidelity, and diversity, producing high-quality synthetic images. The results showed that the hybridized GAN (LS+WS+CGAN) produced lowest D loss of -0.9784, FID of 372.56 and highest SSIM 0.3201 to achieve the best result compare to individual CGAN, LSGAN and WGAN.